Salesforce has released the most accurate speech recognition model anyone can download. SOV-Transcribe-1, unveiled on November 12 by the company's Sovereign Intelligence team, leads the open-weights ranking on Artificial Analysis's speech benchmark with a 2.28% word error rate — ahead of Mistral's Voxtral Small and Nvidia's Nemotron 3.5 Transcribe, and seventh among all 44 models tested, proprietary systems included.

The catch is written into the same model card that reports the score. The model only transcribes finished recordings, covers seven European languages, and needs roughly 53 GB of weights in memory — Salesforce's own serving notes assume a 96 GB GPU. It is a transcriber built for meetings, calls and archives that an organization wants to keep on its own hardware, not for live captions.

The numbers#

Artificial Analysis's AA-WER Index measures word error rate on recorded speech. SOV-Transcribe-1's 2.28% is the best figure any downloadable model has posted, and the margin over its nearest open rivals is not small:

ModelAA-WEROpen weights
SOV-Transcribe-1 (Salesforce)2.28%Yes — Apache 2.0
Voxtral Small (Mistral)2.77%Yes — Apache 2.0
Nemotron 3.5 Transcribe (Nvidia)2.81%Yes

The model also leads on the narrower tests inside the same evaluation — 4.1% word error on Earnings22, 1.5% on VoxPopuli, 1.8% on AA-AgentTalk — and on VoiceCodeBench, a separate benchmark that grades whether a transcript gets names, numbers and entities right, it scores 76.3% task success and 93.4% entity recovery, second of 23 systems overall. Those are Salesforce-reported figures, and the rankings are dated November 12, the day of release. Independent reproduction will be the real test, but the leaderboard itself is third-party.

A secure data center at night, server racks behind glass with a glowing audio waveform on a wall display
Self-hosted transcription keeps audio inside an organization's own infrastructure. Editorial illustration: AI Frontier Post.

What's under the hood#

SOV-Transcribe-1 is a speech-augmented language model, not a standalone speech encoder. A Whisper large-v3 audio encoder turns sound into features, a small projector stacks those features into embeddings, and Google's Gemma 4 26B A4B mixture-of-experts model writes the transcript. The total is 26.6 billion parameters, but only 4.6 billion are active for any given token — the rest still has to sit in memory, which is why the footprint is large and the compute bill is not.

The language-model half is what makes the system more than a transcriber. It takes natural-language instructions alongside the audio: ask for a verbatim transcript or a cleaned-up one, set punctuation and capitalization rules, choose how numbers are written. It also accepts context — a list of names, domain vocabulary, a short description of the topic — to help it recognize difficult words, and it is tuned to get entities right: names, phone numbers, email addresses, the details that matter in dictation and voice applications. A companion draft model, DFlash, proposes tokens for the main model to verify, speeding up decoding roughly two to three times.

Built for the record, not the live call#

The limits matter as much as the score. SOV-Transcribe-1 is non-streaming by design: it transcribes a finished recording, up to 30 minutes at a time, and will not emit words while someone is still speaking. Anyone building live captions or a voice agent that must respond mid-sentence needs a different system.

The language list is seven, all European — English, French, German, Spanish, Italian, Portuguese and Dutch — and English dominates the published benchmarks. The model card also warns of rare repetition loops that need post-processing, and native support in vLLM and Hugging Face Transformers is still being upstreamed; until it lands, users need a plugin for vLLM 0.32.0 or the modeling code shipped in the repository.

A laptop showing a video call with four participants beside a glowing audio waveform
Recorded meetings and calls are the model's home turf. Editorial illustration: AI Frontier Post.

Why a CRM company is shipping speech models#

SOV-Transcribe-1 sits inside Salesforce's Sovereign AI initiative and its Missionforce unit, launched in September 2025 to sell secure, controllable AI to governments and regulated industries. For those customers, an open-weights model is not a philosophical choice: audio that never leaves the building is often a procurement requirement, and a model that can be deployed inside an agency's own environment clears a bar no API can.

That framing explains the design choices — batch transcription of recordings rather than streaming, style control rather than speed, accuracy over languages. It also explains why Salesforce is publishing weights under Apache 2.0 instead of keeping the model behind an API: the customers it wants cannot use an API for this workload anyway. The open-weights leaderboard is the proof of competence; the sovereign pitch is the business.

What to watch#

Three things. First, the plumbing: whether native vLLM and Transformers support lands soon, and how much that lowers the barrier to running a 53 GB model. Second, the response: Mistral and Nvidia both have open transcribers within half a point, and neither is likely to leave the top of the open chart uncontested for long. Third, the sovereign bet itself: whether government and regulated buyers adopt a self-hosted Salesforce transcriber at scale — the outcome that would turn a benchmark win into a product line.

Sources: Hugging Face — SOV-Transcribe-1 model card; Crypto Briefing — Salesforce releases SOV-Transcribe-1, now the top open-weights speech model (Nov. 12, 2026); AI Tech Archive — Open Weights Speech Model: SOV-Transcribe-1 vs Voxtral; ByteWoops — Salesforce's SOV-Transcribe-1 leads open-weight batch speech-to-text. Technical report: arXiv:2611.16919.